61 lines
1.4 KiB
Plaintext
61 lines
1.4 KiB
Plaintext
|
MODELSCOPE_CACHE=./modelscope_cache
|
|||
|
|
|
|||
|
|
MODEL_ID=gemma-4-26B
|
|||
|
|
MODEL_DIR=./models/gemma-4-26B-A4B-it-FP8
|
|||
|
|
PORT=9527
|
|||
|
|
|
|||
|
|
# 指定加载到哪些显卡0,1,2,3,4
|
|||
|
|
CUDA_VISIBLE_DEVICES=0
|
|||
|
|
# 张量并行卡数
|
|||
|
|
TENSOR_PARALLEL_SIZE=1
|
|||
|
|
|
|||
|
|
# 上下文长度
|
|||
|
|
MAX_MODEL_LEN=81920
|
|||
|
|
|
|||
|
|
# 显存占用比例。默认参数0.9，多余显存分配个KV Cache以支持高并发
|
|||
|
|
GPU_MEMORY_UTILIZATION=0.30
|
|||
|
|
|
|||
|
|
# 计算精度
|
|||
|
|
# DTYPE=bfloat16
|
|||
|
|
|
|||
|
|
# KV Cache 精度（auto/fp8）
|
|||
|
|
# KV_CACHE_DTYPE=auto
|
|||
|
|
|
|||
|
|
# 最大并发序列数
|
|||
|
|
MAX_NUM_SEQS=64
|
|||
|
|
|
|||
|
|
# 单批最大 token 数，根据并发和实际上下文需求配置，默认自动分配
|
|||
|
|
MAX_NUM_BATCHED_TOKENSMAX=8192
|
|||
|
|
|
|||
|
|
|
|||
|
|
# 其他运行开关
|
|||
|
|
DISABLE_LOG_REQUESTS=False
|
|||
|
|
ENABLE_LOG_REQUESTS=true
|
|||
|
|
|
|||
|
|
# VLLM运行模式DEBUG\INFO
|
|||
|
|
VLLM_LOGGING_LEVEL=INFO
|
|||
|
|
|
|||
|
|
# Tool calling 配置，需要和模型配套，否则可能出现工具调用失败
|
|||
|
|
ENABLE_AUTO_TOOL_CHOICE=true
|
|||
|
|
TOOL_CALL_PARSER=gemma4
|
|||
|
|
REASONING_PARSER=gemma4
|
|||
|
|
|
|||
|
|
# Gemma专用推理链额外标记开关
|
|||
|
|
DEFAULT_CHAT_TEMPLATE_KWARGS='{"enable_thinking": true}'
|
|||
|
|
|
|||
|
|
# 留空时使用模型自带 chat_template；如需官方工具模板可填绝对路径
|
|||
|
|
# CHAT_TEMPLATE=
|
|||
|
|
|
|||
|
|
# enable-chunked-prefill分块预填空，避免瞬时大量占用内存
|
|||
|
|
# max-num-partial-prefills 1限制同一时刻最多只有 1 个 处于“分块预填充中”的请求，提高稳定性。
|
|||
|
|
|
|||
|
|
TRUST_REMOTE_CODE=true
|
|||
|
|
API_KEY=unis123
|
|||
|
|
|
|||
|
|
# 采样参数
|
|||
|
|
# TEMPERATURE=1.0
|
|||
|
|
# TOP_P=0.95
|
|||
|
|
# TOP_K=64
|
|||
|
|
|
|||
|
|
LOG_DIR=./logs
|